Structural health monitoring of bridges: a model-free ANN-based approach to damage detection

Abstract

As civil engineering structures are growing in dimension and longevity, there is an associated increase in concern regarding the maintenance of such structures. Bridges, in particular, are critical links in today's transportation networks and hence fundamental for the development of society. In this context, the demand for novel damage detection techniques and reliable structural health monitoring systems is currently high. This paper presents a model-free damage detection approach based on machine learning techniques. The method is applied to data on the structural condition of a fictitious railway bridge gathered in a numerical experiment using a three-dimensional finite element model. Data are collected from the dynamic response of the structure, which is simulated in the course of the passage of a train, considering the bridge in healthy and two different damaged scenarios. In the first stage of the proposed method, artificial neural networks are trained with an unsupervised learning approach with input data composed of accelerations gathered on the healthy bridge. Based on the acceleration values at previous instants in time, the networks are able to predict future accelerations. In the second stage, the prediction errors of each network are statistically characterized by a Gaussian process that supports the choice of a damage detection threshold. Subsequent to this, by comparing damage indices with said threshold, it is possible to discriminate between different structural conditions, namely between healthy and damaged. From here and for each damage case scenario, receiver operating characteristic curves that illustrate the trade-off between true and false positives can be obtained. Lastly, based on the Bayes' Theorem, a simplified method for the calculation of the expected total cost of the proposed strategy, as a function of the chosen threshold, is suggested.

An approach to decision‐making analysis for implementation of structural health monitoring in bridges

“…The background of the adopted method for damage detection is proposed in Neves, and explaining it in detail is therefore beyond the scope of this paper. For a deeper insight, the interested reader is referred to the abovementioned work, but nevertheless, the essential points are here described.…”

“…The structure is replicated in healthy and damaged states. The first stage of the proposed method consists in the unsupervised training of Artificial Neural Networks that predict future output acceleration given input data composed of measured accelerations in previous instants. Unsupervised learning means that the algorithm looks at unlabeled input data during the training phase, making it fit for clustering, for example, while supervised learning acts on data labels, making it fit for classification and regression.…”

“…Following the former steps, it is possible to generate a Receiver Operating Characteristic (ROC) curve for the system, based on which the damage detection threshold can be adjusted in order to achieve the system's optimal performance by means of controlling the trade‐off between the two types of errors. The emphasis of the present paper is on this latter part because it was very briefly contemplated in the earlier work of Neves . The authors propose to describe the decision problem concerning the damage detection system setup from a cost optimization standpoint, using tools and concepts such as Bayes' Theorem, decision trees, and EUT.…”

“…The case study concerns the structure addressed in the study of Neves, a simply supported composite railway bridge, which is described with more detail in the referred work.…”

The synthesis of data from instrumented structures and physics-based models via Gaussian processes

“…A multi-output Gaussian process model is also utilised in Zhou and Tang (2018); this differs from the work presented in this paper as it combines simulated data from two physics-based models with different fidelities (accuracies), instead of combining simulation from a physics-based model and observed data. Gaussian processes have also recently been used for other aspects of structural health monitoring in Neves et al (2017); Worden and Cross (2018); Teimouri et al (2017); Fuentes et al. The benefits of the presented work are two-fold: First, through the proposed joint model one obtains aposteriori estimates for the response of the structure, inferred from both observed data and simulations from an analytical physics-based model.…”